Hybrid deep learning and texture-augmented spectral features for sugarcane mapping

Timely identification of sugarcane cultivation is essential for precision agriculture, policy formulation, and sustainable land management. Our study presents a multimodal framework that integrates spectral, morphological, and textural features from Sentinel-2 imagery with hybrid deep learning techniques for sugarcane mapping. A comprehensive feature space combining NDVI, geometrical descriptors, and GLCM-based texture metrics was used to train Random Forest (RF) and Support Vector Machine (SVM) classifiers. A hybrid convolutional neural network (CNN+MLP) architecture further merged visual embeddings from image patches with tabular descriptors. The RF model achieved a macro F1-score of 0.91, a Kappa coefficient of 0.86, and a macro-AUC of 0.94 across agricultural regions in Brazil. DBSCAN clustering was applied to filter noisy labels and improve the quality of the training dataset. A domain-specific confusion cost matrix illustrated the economic impact of misclassification, showing that the RF and CNN+MLP models improved accuracy and financial efficiency by reducing crop identification errors. Although the analysis was limited to São Paulo and Mato Grosso do Sul, the framework showed consistent and robust performance within these regions. Overall, it highlights the benefits of combining morphospectral, textural, and deep features for crop classification, providing a scalable foundation for time-series modeling and real-time land-use analytics. • Sugarcane regions were accurately mapped using Sentinel-2 satellite imagery across different agricultural zones in Brazil. • A hybrid CNN+MLP model combined visual embeddings and tabular features to improve classification. • Morphological, spectral (NDVI), and GLCM-based texture features were fused to improve crop discrimination. • Random Forest and Support Vector Machine classifiers were benchmarked and compared with the deep learning approach. • Unsupervised DBSCAN clustering was applied to detect and remove noisy samples, improving training set quality.